ayameRushia/wav2vec2-large-xls-r-300m-el

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - EL dataset. It achieves the following results on the evaluation set:

Loss: 0.3218
Wer: 0.3095

Training and evaluation data

Evaluation is conducted in Notebook, you can see within the repo "notebook_evaluation_wav2vec2_el.ipynb"

Test WER without LM wer = 31.1294 % cer = 7.9509 %

Test WER using LM wer = 20.7340 % cer = 6.0466 %

How to use eval.py

huggingface-cli login #login to huggingface for getting auth token to access the common voice v8
#running with LM
!python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-el --dataset mozilla-foundation/common_voice_8_0 --config el --split test
# running without LM
!python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-el --dataset mozilla-foundation/common_voice_8_0 --config el --split test --greedy

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 32
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 400
num_epochs: 80.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.3683	8.77	500	3.1280	1.0
1.9915	17.54	1000	0.6600	0.6444
0.6565	26.32	1500	0.4208	0.4486
0.4484	35.09	2000	0.3885	0.4006
0.3573	43.86	2500	0.3548	0.3626
0.3063	52.63	3000	0.3375	0.3430
0.2751	61.4	3500	0.3359	0.3241
0.2511	70.18	4000	0.3222	0.3108
0.2361	78.95	4500	0.3205	0.3084

Framework versions